Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Biologically Inspired Learning Model for Instructed Vision
Authors: Roy Abel, Shimon Ullman
NeurIPS 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our BU-TD model, learned via Counter-Hebb learning, in two settings: 1) unguided visual processing, to show that CH learning is capable of learning vision models 2) guided visual processing, to evaluate the ability of our model to guide the visual process according to instructions. |
| Researcher Affiliation | Academia | Roy Abel Weizmann Institute of Science EMAIL Shimon Ullman Weizmann Institute of Science EMAIL |
| Pseudocode | Yes | Algorithm 1 Counter-Hebb Learning (Section 4.1) and Algorithm 2 Instruction-Based Learning (Section 5) |
| Open Source Code | Yes | The code for reproducing the experiments and creating BU-TD models for guided models is available at https://github.com/royabel/Top-Down-Networks. |
| Open Datasets | Yes | In the unguided experiments, we evaluate the performance of the Counter-Hebb learning on standard image classification benchmarks: MNIST [Le Cun et al., 1998], Fashion-MNIST [Xiao et al., 2017], and CIFAR10 [Krizhevsky et al., 2009]. We followed the same experiments as Bozkurt et al. [2024] and used two-layer fully connected networks, with a hidden layer of size 500 for both MNIST and Fashion-MNIST datasets and size 1,000 for CIFAR10. Further details including the full set of hyperparameters can be found in Appendix A.4.2. |
| Dataset Splits | No | We omitted the validation set, and the hyper-parameters were tuned based solely on the training set. |
| Hardware Specification | Yes | All the experiments were conducted using either NVIDIA RTX 6000 GPU or NVIDIA RTX 8000 GPU. For all experiments but Celeb A, a single NVIDIA RTX 6000 GPU was used, with the experiments utilizing only a fraction of its capacity. In the case of the Celeb A dataset, either a single NVIDIA RTX 8000 GPU or two NVIDIA RTX 6000 GPUs were used. |
| Software Dependencies | No | The paper mentions software components such as 'The standard Adam optimizer [Ruder, 2016]' and the use of 'Res Net-18 [He et al., 2016] architecture (without the final layer) with batch normalization layers [Ioffe and Szegedy, 2015]' (Section 6.2). However, it does not specify version numbers for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA) used in the experiments. |
| Experiment Setup | Yes | We trained for 50 epochs with an exponential learning rate decay with γ = 0.95. The initial learning rate was 10 4, and the batch size 20. |